Data Adaptive Multi-Band Approach for Robust Audio Watermarking Using Adaptive Threshold

(1)

© 2013. K.M. Ibrahim Khalilullah, A K M Wasif Abedin, Mahfuza Khatun, Mohammad Tareq & Md. A.R. Pramanik. This is a research/review paper, distributed under the terms of the Creative Commons Attribution-Noncommercial 3.0 Unported License http://creativecommons.org/licenses/by-nc/3.0/), permitting all non-commercial use, distribution, and reproduction inany medium,

Global Journal of Computer Science and Technology

Graphics & Vision

Volume 13 Issue 3 Version 1.0 Year 2013

Type: Double Blind Peer Reviewed International Research Journal Publisher: Global Journals Inc. (USA)

Online ISSN: 0975-4172 & Print ISSN: 0975-4350

Data Adaptive Multi-Band Approach for Robust Audio

Watermarking using Adaptive Threshold

By K.M. Ibrahim Khalilullah, A K M Wasif Abedin, Mahfuza Khatun,

Mohammad Tareq & Md. A.R. Pramanik

Dhaka International University, Dhaka, Bangladesh

Abstract -

This paper presents a data adaptive digital audio watermarking process with the use of Empirical Mode Decomposition (EMD) and Hilbert Transform (HT). The host audio signal is decomposed into several Intrinsic Mode Functions (IMFs). A binary image or a set of -1 and 1, which is obtained by mapping from standard normal distributed pseudo random numbers, is embedded as secret information into the significant coefficients of the highest energetic IMF that are greater than a specified adaptive threshold. Thus watermarked IMF is less sensitive to common signal processing attack. As a result this method increases more robustness. The experimental results show that the proposed method has good imperceptibility and robustness under common signal processing attacks such as additive noise (in time domain and in frequency domain), low pass filtering, re-sampling, re-quantization, MP3 compression, and sound processing effects such as, delay, a natural sounding reverberation (Schroeder's Reverberator), flanging effect, equalization effect, etc. A Bit Error calculation equation is provided to measure the accuracy of the extracted watermark.

GJCST-F Classification:

I.4.5

Data Adaptive Multi-Band Approach for Robust Audio Watermarking using Adaptive Threshold

Strictly as per the compliance and regulations of:

Keywords :

audio watermarking, empirical mode decomposition, hilbert transform, wavelet transform, adaptive

threshold, binary image.

(2)

Data Adaptive Multi-Band Approach for Robust

Audio Watermarking using Adaptive Threshold

Abstract - This paper presents a data adaptive digital audio watermarking process with the use of Empirical Mode Decomposition (EMD) and Hilbert Transform (HT). The host audio signal is decomposed into several Intrinsic Mode Functions (IMFs). A binary image or a set of -1 and 1, which is obtained by mapping from standard normal distributed pseudo random numbers, is embedded as secret information into the significant coefficients of the highest energetic IMF that are greater than a specified adaptive threshold. Thus watermarked IMF is less sensitive to common signal processing attack. As a result this method increases more robustness. The experimental results show that the proposed method has good imperceptibility and robustness under common signal processing attacks such as additive noise (in time domain and in frequency domain), low pass filtering, re -sampling, re-quantization, MP3 compression, and sound processing effects such as, delay, a natural sounding reverberation (Schroeder's Reverberator), flanging effect, equalization effect, etc. A Bit Error calculation equation is provided to measure the accuracy of the extracted watermark.

Keywords : audio watermarking, empirical mode decomposition, hilbert transform, wavelet transform, adaptive threshold, binary image.

I.

I

ntroduction

n any given application, practitioners had already recognized two different technologies for multimedia data protection: encryption and digital watermarking [1, 2]. However, a digital watermark is complementary approach than cryptographic processes. The availability

of digital multimediacontents has been grown rapidly in

the recent years. Today, digital media documents can be distributed via the World Wide Web to a large number of people without much effort and money. Unlike traditional analog copying, with which the quality of the duplicated content is degraded, digital tools can easily produce large amount of perfect copies of digital documents in a short period.

This ease of digital multimedia distribution over

the Internet, together with the possibility of unlimited

duplication of this data, threatens the intellectual property (IP) rights more than ever. Therefore, there is a

strong need for security services in order to keep

ownership for the document owner and reliable to the customer. Watermarking plays an important role for that purposes. In a digital watermarking process, typically, a

pattern or sequence of bits is embedded into a digital

image, an audio or video file such that the embedded pattern can be detected or extracted later to make an assertion about the object. Some proposed or actual watermarking applications are [3]: broadcast monitoring, owner identification, proof of ownership, transaction tracking, content authentication and tampering detection, copy control, and device control, fingerprinting, information carrier.

Embedding secret information into audio

sequences is a more tedious task than image

watermarking, due to dynamic preeminence of the human auditory system (HAS) over human visual system (HVS). The main confront in digital audio watermarking and steganography is that if the perceptual transparency parameter is predetermined, the design of a watermark system cannot obtain high robustness and a high watermark data rate at the same time. We have addressed this challenging situation with the use of EMD specially developed for analyzing non-linear and non-stationary signals [4].

In the wavelet transform domain [5, 6], the watermark is embedded in high frequency band or in low frequency band; therefore watermark can’t be embedded in the whole transform domain. In EMD-based method, we can add the watermark into the whole transform domain. The EMD method provides perfect time-frequency localization properties due to its instantaneous frequency than discrete wavelet transform (DWT) domain and leads to implicit audio-visual masking [7]. This decomposition method is adaptive, and highly efficient for perceptual transparency (inaudibility) and robustness. That is HAS is perfectly met too. In the previous EMD-based method [8], the watermark is added into the whole coefficients of the highest energetic IMF, therefore the watermarked IMF is sensitive to common signal processing attacks than this proposed method. This paper presents a method, which

I

Global Journal of Computer Science and Technology Volume XIII Issue III Version I

25 (DDDDDDDD ) Year 013 2 F

Author α ¥ : Lecturer, Dept. of Computer Science & Engineering, Dhaka International University, Dhaka, Bangladesh.

E-mails : [email protected], [email protected]

Authorσ:B.Sc Engineering from Dhaka International University major in Computer Science & Engineering. E-mail : [email protected]

Author ρ Ѡ : Lecturer, Dept. of Electrical, Electronics & Telecom-munication Engineering, Dhaka International University, Dhaka, Bangladesh. E-mails : [email protected],

[email protected]

K.M. Ibrahim Khalilullahα_{, A K M Wasif Abedin}σ_{, Mahfuza Khatun}ρ_{, Mohammad Tareq}Ѡ

(3)

II.

P

roposed

A

udio

W

atermarking

T

echnique

The Empirical Mode Decomposition is a new

method for analyzing nonlinear andnon-stationary data.

This method is adaptive, and, therefore, highly efficient. Since this decomposition method is based on the local characteristic time scale of the data, it is applicable to nonlinear and non-stationary processes. Using this method any complicated data set can be decomposed into a finite and often small number of ‘Intrinsic Mode Functions (IMF)’. An Intrinsic mode Function (IMF) is a function that satisfies two conditions [9]: (i) in the whole data set, the number of extrema and the number of zero crossings must either equal or differ at most by one; and (ii) at any point, the mean value of the envelope defined by the local maxima and the envelope defined by the local minima is zero. With the Hilbert transform, the ‘Intrinsic Mode Functions’ yield instantaneous frequencies as functions of time that provide sharp identifications of imbedded structures.

We have extended the previous idea by using EMD and embedding watermark into the filtered coefficients of the highest energetic IMF for increasing robustness against different signal processing attacks. In the proposed method, at first we have divided the host signal into a number of frames. Then each frame is decomposed into a small number of intrinsic mode functions (IMFs) that acknowledge well-behaved Hilbert transforms. This method is applicable to nonlinear and non-stationary process because of the local characteristic time scale of the data. However, after calculating energy of the every IMF, we have picked an IMF contained highest energy. The following Fig. 1 shows the energy distribution of the IMFs for the first frame of an audio signal:

Figure1: Energy distribution for IMFs of the firstframe From the energy distribution of IMFs, we can

select the 6th _{IMF because of the highest energetic}

value. A scaled version of the pseudo random

watermark sequence

w

[ ]

n

, n is the size of the

watermark which is equal to the number of frame, added

to the filtered coefficients of the selected IMFs. Onlyone

bit of

w

[ ]

n

is added to the only one frame. The ith _{bit of}

[ ]

n

w

is added to the selected IMF (e.g. 6th_{IMF) of the i}th

)

(

)

(

)

(

l

d

l

w

i

d

_m

′

=

_mi

+

α

(1) Where, i=1 for the first frame, i=2 for the

second frame and so on.

l

runs over all coefficients of

the IMF > T (Adaptive threshold),

d

_mi

(l

)

is the filtered coefficient of mth_{IMF of i}th_frame,

_w

_(i

₎

_{is the i}th_{bit of the}

watermark, and

d

_m

′

(l

)

is the watermarked coefficient of

the selected IMF. The watermarked frame is generated by simply summing all IMF of that frame. The

watermarked signal

x′

[ ]

t

is generated after embedding

watermarks to all frames in this way. The scaling factor

α

is chosen to be as high as possible while remaining

inaudible. Here

α

is taken as 0.02. The threshold value

(T) is calculatedby this equation (2):

2 /

)

_

(

Max

Val

Min

Val

T

=

+

(2)

Where

Max _

Val

is the maximum coefficient

value and

Min _

Val

is the minimum coefficient value of

the selected IMF.

Since, EMD is empirical, there is no reason the decomposition will result in same (or even similar) IMFs when the signal is modified (or even the watermark added). If it does not result in the same IMFs, the entire

scheme fails. Therefore, in the detection process, the

original signal

x

[ ]

t

and watermarked signal

x′

t

H

[ ]

x

[ ]

t

/

α

w

_i

′

[ ]

t

is the extracted watermark of the ith

frame. The resulted watermark bit is calculated by averaging over the number of embedded watermark to the frame. In this calculation, two different techniques are applied to two different watermarks.

For the sequence of 1 and -1, the watermark bit is a +1 If the average is greater than zero and if the Data Adaptive Multi-Band Approach for Robust Audio Watermarking using Adaptive Threshold

2 ( DDDD ) Year 013 2 26 F

frame by using the following additive embedding formula (1):

the highest energetic IMF using adaptive threshold. Since adaptive threshold is used to add watermark into the significant IMF, our method is much more resistant to common signal processing manipulation when compared with other methods.

(4)

existence of watermark. The similarity should be increased or decreased according to the amount of the inserted watermark, as well as to the degree of the

attack. To calculate the detector response value (

S

dr),

dr

means detector response, the proposed method

utilizes the vector projection as defined in the following equation (4):

[ ] [ ]

n

w

n

w

[ ] [ ]

n

w

n

w

S

_dr

=

⋅

′′

/

′′

⋅

′′

(4)

Where,

w

[ ]

n

is the original watermark or fake

watermark, and

w ′′

[ ]

n

is the resulted watermark. We

also calculated the normalized correlation (

(5)

When the detector response value of the original watermark and extracted watermark is greater than other detector response values of the fake watermarks and resulted watermark, then we can conclude that the watermark is present. Fig. 2 shows a block diagram of the proposed method for a single frame.

(a)

(b)

Figure 2 : The proposed method: (a) Watermark embedding and (b) Watermark detection

III.

E

xperimental

R

esults and

D

iscussion

To test the algorithm we have used eight audio music clips (44 kHz, 705 kpbs), which has been collected from Fruity Loops software. An original audio signal (drum music), watermarked audio signal, and the extracted watermark signal are shown in Fig. 3. After converting the original and watermarked audio signal into HT domain, the extracted watermark is calculated using equation (3).

(a)

(b)

than zero; if the average is less than equal zero the corresponding watermark bit is a zero. Finally we get the

resulted watermark

w ′′

[ ]

n

after calculating the above

operation for all frames.

The similarity between the original watermark and resulted watermark is calculated to detect the average is less than zero the corresponding watermark bit is a –1. If we use binary number as a watermark , then the watermark bit is a 1 if the average is greater

(5)

(c)

Figure 3 : (a) Original audio signal, (b) Watermarked audio signal, (c) Extracted Watermark

The detector response values (

S

_dr) on the

y-axis and randomly generated watermark on the x-y-axis against three attacks as example for the piano music and drum music are shown in the following figures (from Fig. 4 to Fig. 6). The original audio signal is watermarked with a seed of 100. Therefore, in each figure, the detector response values with the original watermark are located at 100 on the x-axis. Therefore, at this position the detector response values are greater than any other values.

(a) (b)

Figure 4 : Detector responses for Gaussian noise attack. (a) Piano music with = 0.96, SNR= -7.26 db and

(b) Drum music with =0.9, SNR= -15.104db

(a) (b)

Figure 5 : Detector responses for Flanging effect (a) Piano music with =1, SNR= 6.02db and (b) Drum

music with =1, SNR= 6.05db

(a) (b)

Figure 6 : Detector responses for MP3 compression

(a) Piano music with

ψ

=1 and (b) Drum music with

ψ

=0.9, 56 kbps

The performance against Gaussian noise addition is calculated to illustrate the robustness of the proposed algorithm as:

F W p

V

R

=

−

(6)

Where,

V

_W is the similarity measurement value

between original watermark and resulted watermark extracted from the attacked watermarked signal and

F

V

is the highest similarity measurement value between

false watermark and resulted watermark extracted from the attacked watermarked signal. The changes of the robustness performance for drum music are shown in Fig. 7:

Figure 7 : Robustness performance values against different noise levels (db)

If

R

p

>

0

, then watermark is detected

correctly. The highest value of

R

_p represents maximum

performance. The proposed method works for -26db

also. We have also calculated Bit Error (

β

) to examine

extracted watermark by the following equation (7):

2 ( DDDD ) Year 013 2 28 F

(6)

( )

L

j

w

j

w

j

2 ∑

−

′′

=

β

(7)

Where, L is the length of the watermark i.e. the

number of bit of the watermark. The minimum value of

Table 1 : Experimental results of different attacks for various types of audio signals

Music Clip

Drum C_HC

results of the eight music clips are shown in Table-I.

The average bit errors of the eight music clips for different Gaussian noise addition are sown in Fig. 8:

Figure 8 : Average bit-error against different level of

IV.

C

onclusion

This paper proposed an audio watermarking schema using the principle advantages of Empirical Mode Decomposition. A multiband approach is presented here for digital audio watermarking. The data adaptive time varying filtering is implemented for sub band decomposition of the host audio signal using EMD. Traditional frequency domain filtering techniques (i.e. Fourier and wavelet) depend on the priori basis functions and hence the signal is fitted with those bases. In the proposed method, the decomposition is fully data dependent and without the use of any basis function. The EMD based decomposition is a complete decomposition i.e. the original signal can easily be obtained by simply summing up the individual IMFs and

(7)

hence the method performs better for audio watermarking. The efficiency of the proposed algorithm has been shown by a series of experiment. In this method we can embed a binary logo image as a watermark also.

R

eferences

R

éférences

R

eferencias

1. M. Eskicioglu & E. J. Delp(April 2001). Overview of

Multimedia Content Protection in Consumer

Electronics Devices. Signal processing: Image

Communication,16(7),pp:681-699.

2. A. M. Eskicioglu &E.J. Delp(April 2003).Security of

Digital Entertainment Content From Creation to

Consumption. Signal Processing : Image

Communication, Special Issue on Image Security,

118 (4), pp.237-262.

3. I. J. Cox, M. L. Miller, and J. A. Bloom (2002).

Digital Watermarking.Morgan Kaufmann Publishers.

4. Huang, N. E. et. al. (1998). The empirical mode

decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis.

Proc. Roy. Soc. London A,Vol. 454, pp. 903-995.

Algorithm. International Journal of Information

Technology, China,Vol. 11 No.12.

7. LingFei Liang, ZiLiang Ping (2008). Information

Hiding Based on Empirical Mode Decomposition. 2008 IEEE Pacific-Asia Workshop on Computational

Intelligence and Industrial Application,paciia, vol. 1,

pp.561-565.

8. A. N. K. Zaman, K. M. I. Khalilullah, M. W. Islam and

M. K. I. Molla (May, 2010). A robust digital audio

watermarking algorithm using empirical mode

decomposition. Proc. of 23rd Canadian Conf. on

Electrical and Computer Engineering, (CCECE'10).

9. _{Foo S, Yeo T & Huang D. (2001). An adaptive audio}

watermarking system. In: Proc. IEEE Region 10 International Conference on Electrical and Electronic Technology, Phuket Island-Langkawi Island, Singapore, p 509–513.

2 ( DDDD ) Year 013 2 30 F

Data Adaptive Multi-Band Approach for Robust Audio Watermarking using Adaptive Threshold

5. Xianghong Tang, Yamei Niu, Hengli Yue, Zhongke Yin (2005). A Digital Audio Watermark Embedding 6. Martinez-Noriega, R. et. al. (July 8-11, 2007). Audio Watermarking Based on Wavelet Transform and

Quantization Index Modulation. The 22nd

International Technical Conference on Circuits/-Systems, Computers and Communications, ITC-CSCC 2007, pp. 133-134, Busan, Korea.